From Prompt Tennis to Maturing Agents: Why Context Switching Feels Like Cognitive Decline
A discussion caught my attention recently on the corporate finance subreddit. A financial analyst at a large company asked whether using artificial intelligence was making them dumb. After handing their financial models and monthly presentation decks over to Claude, they noticed an uneasy feeling. Their mental arithmetic felt sluggish. They worried they were turning into a passive reviewer rather than an active builder.
I recognized that feeling immediately. It crosses my mind most days. When a prompt can complete a task in the background while I focus on something else, taking advantage of it makes practical sense. We build automated workflows because we want our time back. In an ideal setup, I want to trigger an agent in the morning and head out for a run. I want to spend quiet time in the garden, then return to review the finished output with confidence.
In reality, an open chat window rarely delivers that calm routine. Most people end up trapped in a loop of prompt tennis, and the mental fatigue that follows has very little to do with losing basic intelligence.
The Hidden Tax of Prompt Tennis
When you work inside a standard conversational interface, you seldom send one instruction and step away from your desk. You paste a query and wait. Once the output lands on your screen, you spot an omission and type another correction.
I do this frequently. Even when working on technical articles or architecture drafts, I get caught prompting back and forth to make the text sound right. One prompt adjusts the voice, and the next one fixes a missing detail.
Steering a model through micro-adjustments creates continuous friction. The context switching is much more expensive than people realize.
When you write code or analyze a spreadsheet directly, your brain holds a single mental model in working memory. You stay inside that operational flow. During prompt ping-pong, your working memory reloads the problem context with each new generation. You read an unfamiliar draft. You check the facts, and then you figure out how to steer the prompt next.
That operational rhythm mimics the day of an engineering manager interrupted by an eager junior colleague throughout the afternoon. You write very little yourself. Still, by four o’clock, you feel drained. People interpret that fatigue as cognitive decline, but the drain comes from rapid context switching.
Reviewer Fatigue and the Trap of False Greens
Prompt tennis also brings a serious engineering risk into production workflows. It lowers your defenses against false greens.
A false green happens when an agent completes a task and reports a clean status, but introduces a subtle logical error beneath a polished surface. In financial models, that might mean a shifted date column. In backend software, it is often a quietly swallowed database error or an unindexed query. I explored this exact failure pattern and how to benchmark against it in Evaluating Prompts When Decimals Cost Thousands.
Catching those subtle discrepancies requires deep focus. When you review an output after ten consecutive back-and-forth prompts, your attention is fragmented. You skim the surface. The syntax looks valid, so you approve the result. You end up maintaining code or documentation that you did not write and only half-inspected.
Interactive chat windows make delegating work easy, but babysitting the output keeps your attention chained to the screen.
Shifting from Ad-Hoc Prompts to Maturing Agents
Breaking out of that reactive cycle requires changing how we structure automated work. Instead of treating each task as an ad-hoc conversation in a shared chat window, we should start building dedicated agents designed for specific jobs.
A dedicated agent needs time to mature. When an agent is brand new, it behaves like an apprentice joining a team. A lot of back and forth is needed at first, and that initial setup occupies substantial mental capacity. You have to define strict input schemas and establish what a successful run actually looks like.
Open chat interfaces demand continuous human presence, while dedicated agents compound reliability through iterative constraints.
- Continuous Context Switching: Human attention reloads state on each prompt cycle.
- Reviewer Exhaustion: Fragmented attention leads to missed logic errors and false greens.
- Disposable Context: Corrections stay trapped in single chat threads without compounding.
- Upfront Investment: Concentrated mental capacity spent on strict typing and invariants.
- Retrospective Capture: Discovered edge cases turn into automated regression tests.
- Asynchronous Peace: Work runs independently in the background with confident final review.
That upfront investment pays compounding dividends over time. The maturation process moves across three clear phases:
First comes the exploratory phase. You test the agent against messy real-world inputs, observe where it hallucinates or misinterprets instructions, and correct it directly. You note the exact failure modes.
Next comes the stabilization phase. You take those observed failure modes and turn them into permanent system rules or automated regression checks. Instead of re-explaining the same rule in tomorrow’s chat window, you encode the requirement into the agent’s harness, similar to the retrospective feedback loop described in Closing the Feedback Loop: Reducing AI Agent Context Bloat Through Automated Self-Retrospectives. You eliminate the need for manual steering.
Finally comes the autonomous phase. Once the agent is mature, you can let it run for longer stretches in the background with genuine confidence. You trigger the job and step away.
The start is often hard, but the long-term payoff is real.
Earning Autonomy to Reclaim Your Schedule
The uncomfortable feeling of losing your sharpness is a diagnostic signal. It warns you that your current workflow has traded deep thinking for fragmented supervision.
Using prompts to speed up mundane work makes complete sense. Letting interactive chat tabs dominate your day does not. If a task repeats across your week, stop playing prompt tennis in an open window. Treat the problem as an engineering challenge. Build a single-purpose agent with clean state separation like CQRS for LLM Agents, accept the initial mental effort required to train its constraints, and harden its guardrails through postmortems.
When your agents earn true operational maturity, you stop babysitting tokens. You get back the quiet time to focus on hard architectural trade-offs, or simply spend quiet time in the garden.
Extra Resources
- Subreddit Discussion: Anyone else feel like AI is making you dumb?
- Gloria Mark, Attention Span: A Groundbreaking Way to Restore Balance, Happiness and Productivity (Hanover Square Press, 2023)
- Cal Newport, Deep Work: Rules for Focused Success in a Distracted World (Grand Central Publishing, 2016)
Stuck playing prompt tennis or trying to build autonomous agents?
Let's grab a coffee and talk about moving from interactive chat sessions to reliable, maturing background agents that free up your schedule.
Grab a CoffeeTransparency Disclosure: In compliance with transparency guidelines for AI-assisted content under EU policy, please note that this article was co-authored by Savvas Alexandrou and Gemini 3.8 Flash.
